Re: [PATCH 0/9] cpufreq: tegra194: add Tegra264 support

From: Sumit Gupta

Date: Tue Nov 17 2026 - 04:52:11 EST


Hi Zhongqiu,

Thank you for the review.


On 13/11/26 11:56, Zhongqiu Han wrote:
External email: Use caution opening links or attachments


Hi Sumit,

On 11/5/2026 2:15 AM, Sumit Gupta wrote:
This series adds cpufreq support for Tegra264, which joins the SoCs
already supported by the tegra194-cpufreq driver.

Tegra264 keeps the existing model, where BPMP provides the NDIV limits.
The driver derives the current frequency from the core clock and
reference counters.

The main differences are:
  - Frequency read: through system registers, as on Tegra194, but from
    the 64-bit architected AMU counters. The reference counter ticks at
    the system counter rate (CNTFRQ).

  - Frequency request: through MMIO, as on Tegra234. The driver writes
    NDIV to a core_freq_req register in the CCPLEX register space. Each
    register controls a cluster-clock unit shared by two single-CPU
    clusters, so both CPUs form one cpufreq policy.

The patches fall into four groups:
  - Fix (patch 1): read each CPU's MPIDR from a CPU-hotplug callback
    instead of at probe. A CPU that is offline at probe then also gets
    valid cluster and core IDs.

  - Preparation (patches 2 to 5): cleanups, mostly of the counter code,
    to make room for a 64-bit counter and for an MMIO mapping without
    ACTMON counters.

  - Improvement (patches 6 and 7): group CPUs by the generic
    "performance-domains" property, which cluster-based grouping cannot
    express on Tegra264. Device trees without it keep the existing
    grouping.

  - Tegra264 support (patches 8 and 9): the driver support, followed by
    the device tree changes it relies on.

Patch 9 depends on "arm64: tegra: Fix address translation on Tegra264"
from Thierry's series [1]. Without it, the ccplex@30000000 node does not
translate to the CCPLEX registers.


I have some inline comments. Please take a look.

Sashiko also has some feedback that may be worth checking:

https://sashiko.dev/#/patchset/20261104181521.3003578-1-sumitg%40nvidia.com


I will address the findings on patches 1 and 7 in v2, with checks in
tegra194_get_speed(). The issue in patch 9 comes from applying it
without its dependency.


P1 and P8 don't apply on v7.4-rc2 / linux-next due to 4dd10fbbeb9c
(devm_kmalloc_obj conversion). Could you rebase?


Will do in v2.

Thanks,
Sumit



Sumit Gupta (9):
   cpufreq: tegra194: move MPIDR read to CPU hotplug callback
   cpufreq: tegra194: remove duplicate counter sampling loops
   cpufreq: tegra194: add tegra_counter_delta helper
   cpufreq: tegra194: add per-SoC counter mask
   cpufreq: tegra194: split MMIO mapping from ACTMON offset
   dt-bindings: arm: nvidia: Document the Tegra264 CCPLEX cluster
   cpufreq: tegra194: derive frequency domain from device tree
   cpufreq: tegra194: Add support for Tegra264
   arm64: tegra: Add CCPLEX cluster node for Tegra264

  .../tegra/nvidia,tegra-ccplex-cluster.yaml    |  27 +-
  arch/arm64/boot/dts/nvidia/tegra264.dtsi      |  21 +
  drivers/cpufreq/Kconfig.arm                   |   4 +-
  drivers/cpufreq/tegra194-cpufreq.c            | 422 +++++++++++++-----
  4 files changed, 357 insertions(+), 117 deletions(-)

[1] https://lore.kernel.org/linux-tegra/20261030141101.845044-1- thierry.reding@xxxxxxxxxx/



--
Thx and BRs,
Zhongqiu Han